Unsupervised Machine Learning for the Classification of Astrophysical X-ray Sources

TL;DR

This paper tackles the challenge of classifying Chandra X-ray sources in CSC with limited labeled data by adopting an unsupervised cluster-then-label framework. It uses Gaussian Mixture Models to cluster per-detection X-ray features, links clusters to known classes via SIMBAD and Mahalanobis-distance-based scoring, and outputs probabilistic class assignments for 8,756 sources across 14,507 detections, with master classifications computed through hard and soft voting. The approach habilitates robust identification of young stellar objects and differentiates large accretors (AGN/QSO/Seyfert) from small accretors (X-ray binaries), showing results consistent with the unified AGN model and X-ray variability physics. The work delivers a reproducible methodology with public code and a Streamlit-based playground, enabling broader adoption for upcoming all-sky X-ray surveys and catalogs.

Abstract

The automatic classification of X-ray detections is a necessary step in extracting astrophysical information from compiled catalogs of astrophysical sources. Classification is useful for the study of individual objects, statistics for population studies, as well as for anomaly detection, i.e., the identification of new unexplored phenomena, including transients and spectrally extreme sources. Despite the importance of this task, classification remains challenging in X-ray astronomy due to the lack of optical counterparts and representative training sets. We develop an alternative methodology that employs an unsupervised machine learning approach to provide probabilistic classes to Chandra Source Catalog sources with a limited number of labeled sources, and without ancillary information from optical and infrared catalogs. We provide a catalog of probabilistic classes for 8,756 sources, comprising a total of 14,507 detections, and demonstrate the success of the method at identifying emission from young stellar objects, as well as distinguishing between small-scale and large-scale compact accretors with a significant level of confidence. We investigate the consistency between the distribution of features among classified objects and well-established astrophysical hypotheses such as the unified AGN model. This provides interpretability to the probabilistic classifier. Code and tables are available publicly through GitHub. We provide a web playground for readers to explore our final classification at https://umlcaxs-playground.streamlit.app.

Figures (10)

• Figure 1: The Bayesian Information Criterion (BIC) (top) and the gradient of the BIC (bottom) as a function of the number of components $K$. The number of components ranges from $2$ to $20$. The gray region delineates the confidence interval as determined by the standard deviation of each $K$'s iteration results. The BIC function is smoothly decreasing, while its gradient shows a constant behaviour for values greater than $K=6$. The red dashed vertical line highlights the function point at $K=6$, which is the number of components that shows a better configuration in this technique.
• Figure 2: Source detections in galactic coordinates over a Mollweide projection, discriminated by their assigned cluster in colors and markers. We see a trend of extragalactic or galactic for points in particular clusters.
• Figure 3: hard_hs vs. hard_hm scatter plot for each cluster, with var_prob_h as a color dimension. Notice how some clusters clearly demonstrate multidimensional correlations, predominantly influenced by highly variable or low variable source detections.
• Figure 4: True vs. Predicted confusion matrix for the benchmark set using individual detections only, normalized by row. For each true class the proportion of source detections of that class assigned to each of the possible labels is shown.
• Figure 5: True vs. Predicted confusion matrix for the benchmark set, normalized by row. Classes were replaced by the aggregated classes AGN, Seyfert, QSO, YSO. The matrix displays the proportion of source detections in a specific class that were correctly classified or misclassified. The data used in this plot is not a final result of our pipeline.